[PATCH v6 3/3] blk-cgroup: Optimize blkcg_rstat_flush()

[Waiman Long]

For a system with many CPUs and block devices, the time to do
blkcg_rstat_flush() from cgroup_rstat_flush() can be rather long. It
can be especially problematic as interrupt is disabled during the flush.
It was reported that it might take seconds to complete in some extreme
cases leading to hard lockup messages.

As it is likely that not all the percpu blkg_iostat_set's has been
updated since the last flush, those stale blkg_iostat_set's don't need
to be flushed in this case. This patch optimizes blkcg_rstat_flush()
by keeping a lockless list of recently updated blkg_iostat_set's in a
newly added percpu blkcg->lhead pointer.

The blkg_iostat_set is added to the lockless list on the update side
in blk_cgroup_bio_start(). It is removed from the lockless list when
flushed in blkcg_rstat_flush(). Due to racing, it is possible that
blk_iostat_set's in the lockless list may have no new IO stats to be
flushed. To protect against destruction of blkg, a percpu reference is
gotten when putting into the lockless list and put back when removed.

A blkg_iostat_set can determine if it is in a lockless list by checking
the content of its lnode.next pointer which will be non-NULL when in
a lockless list. This requires the presence of a special llist_last
sentinel node to be put at the end of the lockless list.

When booting up an instrumented test kernel with this patch on a
2-socket 96-thread system with cgroup v2, out of the 2051 calls to
cgroup_rstat_flush() after bootup, 1788 of the calls were exited
immediately because of empty lockless list. After an all-cpu kernel
build, the ratio became 6295424/6340513. That was more than 99%.

Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
Acked-by: Tejun Heo <tj@xxxxxxxxxx>
---
 block/blk-cgroup.c | 100 ++++++++++++++++++++++++++++++++++++++++++---
 block/blk-cgroup.h |   9 ++++
 2 files changed, 103 insertions(+), 6 deletions(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 9021f75fc752..963a779c4cab 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c
@@ -59,6 +59,71 @@ static struct workqueue_struct *blkcg_punt_bio_wq;
 
 #define BLKG_DESTROY_BATCH_SIZE  64
 
+/*
+ * Lockless lists for tracking IO stats update
+ *
+ * New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg).
+ * There are multiple blkg's (one for each block device) attached to each
+ * blkcg. The rstat code keeps track of which cpu has IO stats updated,
+ * but it doesn't know which blkg has the updated stats. If there are many
+ * block devices in a system, the cost of iterating all the blkg's to flush
+ * out the IO stats can be high. To reduce such overhead, a set of percpu
+ * lockless lists (lhead) per blkcg are used to track the set of recently
+ * updated iostat_cpu's since the last flush. An iostat_cpu will be put
+ * onto the lockless list on the update side [blk_cgroup_bio_start()] if
+ * not there yet and then removed when being flushed [blkcg_rstat_flush()].
+ * References to blkg are gotten and then put back in the process to
+ * protect against blkg removal.
+ *
+ * lnode.next of the last entry in a lockless list is NULL. To enable us to
+ * use lnode.next as a boolean flag to indicate its presence in a lockless
+ * list, we have to make it non-NULL for all. This is done by using a
+ * sentinel node at the end of the lockless list. All the percpu lhead's
+ * are initialized to point to that sentinel node as being empty.
+ */
+static struct llist_node llist_last;
+
+static bool blkcg_llist_empty(struct llist_head *lhead)
+{
+	return lhead->first == &llist_last;
+}
+
+static void init_blkcg_llists(struct blkcg *blkcg)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu)
+		per_cpu_ptr(blkcg->lhead, cpu)->first = &llist_last;
+}
+
+static struct llist_node *fetch_delete_blkcg_llist(struct llist_head *lhead)
+{
+	return xchg(&lhead->first, &llist_last);
+}
+
+static struct llist_node *fetch_delete_lnode_next(struct llist_node *lnode)
+{
+	struct llist_node *next = READ_ONCE(lnode->next);
+	struct blkcg_gq *blkg = llist_entry(lnode, struct blkg_iostat_set,
+					    lnode)->blkg;
+
+	WRITE_ONCE(lnode->next, NULL);
+	percpu_ref_put(&blkg->refcnt);
+	return next;
+}
+
+/*
+ * The retrieved blkg_iostat_set is immediately marked as not in the
+ * lockless list by clearing its node->next pointer. It could be put
+ * back into the list by a parallel update before the iostat's are
+ * finally flushed including probably the new update.
+ */
+#define blkcg_llist_for_each_entry_safe(pos, node, nxt)			\
+	for (; (node != &llist_last) &&					\
+	       (pos = llist_entry(node, struct blkg_iostat_set, lnode),	\
+		nxt = fetch_delete_lnode_next(node), true);		\
+		node = nxt)
+
 /**
  * blkcg_css - find the current css
  *
@@ -236,8 +301,10 @@ static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
 	blkg->blkcg = blkcg;
 
 	u64_stats_init(&blkg->iostat.sync);
-	for_each_possible_cpu(cpu)
+	for_each_possible_cpu(cpu) {
 		u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
+		per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
+	}
 
 	for (i = 0; i < BLKCG_MAX_POLS; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
@@ -852,17 +919,23 @@ static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
 {
 	struct blkcg *blkcg = css_to_blkcg(css);
-	struct blkcg_gq *blkg;
+	struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
+	struct llist_node *lnode, *lnext;
+	struct blkg_iostat_set *bisc;
 
 	/* Root-level stats are sourced from system-wide IO stats */
 	if (!cgroup_parent(css->cgroup))
 		return;
 
+	if (blkcg_llist_empty(lhead))
+		return;
+
 	rcu_read_lock();
 
-	hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
+	lnode = fetch_delete_blkcg_llist(lhead);
+	blkcg_llist_for_each_entry_safe(bisc, lnode, lnext) {
+		struct blkcg_gq *blkg = bisc->blkg;
 		struct blkcg_gq *parent = blkg->parent;
-		struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
 		struct blkg_iostat cur, delta;
 		unsigned long flags;
 		unsigned int seq;
@@ -1170,6 +1243,7 @@ static void blkcg_css_free(struct cgroup_subsys_state *css)
 
 	mutex_unlock(&blkcg_pol_mutex);
 
+	free_percpu(blkcg->lhead);
 	kfree(blkcg);
 }
 
@@ -1189,6 +1263,11 @@ blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
 			goto unlock;
 	}
 
+	blkcg->lhead = alloc_percpu_gfp(struct llist_head, GFP_KERNEL);
+	if (!blkcg->lhead)
+		goto free_blkcg;
+	init_blkcg_llists(blkcg);
+
 	for (i = 0; i < BLKCG_MAX_POLS ; i++) {
 		struct blkcg_policy *pol = blkcg_policy[i];
 		struct blkcg_policy_data *cpd;
@@ -1229,7 +1308,8 @@ blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
 	for (i--; i >= 0; i--)
 		if (blkcg->cpd[i])
 			blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
-
+	free_percpu(blkcg->lhead);
+free_blkcg:
 	if (blkcg != &blkcg_root)
 		kfree(blkcg);
 unlock:
@@ -1993,6 +2073,7 @@ static int blk_cgroup_io_type(struct bio *bio)
 
 void blk_cgroup_bio_start(struct bio *bio)
 {
+	struct blkcg *blkcg = bio->bi_blkg->blkcg;
 	int rwd = blk_cgroup_io_type(bio), cpu;
 	struct blkg_iostat_set *bis;
 	unsigned long flags;
@@ -2011,9 +2092,16 @@ void blk_cgroup_bio_start(struct bio *bio)
 	}
 	bis->cur.ios[rwd]++;
 
+	if (!READ_ONCE(bis->lnode.next)) {
+		struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
+
+		llist_add(&bis->lnode, lhead);
+		percpu_ref_get(&bis->blkg->refcnt);
+	}
+
 	u64_stats_update_end_irqrestore(&bis->sync, flags);
 	if (cgroup_subsys_on_dfl(io_cgrp_subsys))
-		cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
+		cgroup_rstat_updated(blkcg->css.cgroup, cpu);
 	put_cpu();
 }
 
diff --git a/block/blk-cgroup.h b/block/blk-cgroup.h
index d4de0a35e066..2c36362a332e 100644
--- a/block/blk-cgroup.h
+++ b/block/blk-cgroup.h
@@ -18,6 +18,7 @@
 #include <linux/cgroup.h>
 #include <linux/kthread.h>
 #include <linux/blk-mq.h>
+#include <linux/llist.h>
 
 struct blkcg_gq;
 struct blkg_policy_data;
@@ -43,6 +44,8 @@ struct blkg_iostat {
 
 struct blkg_iostat_set {
 	struct u64_stats_sync		sync;
+	struct llist_node		lnode;
+	struct blkcg_gq		       *blkg;
 	struct blkg_iostat		cur;
 	struct blkg_iostat		last;
 };
@@ -97,6 +100,12 @@ struct blkcg {
 	struct blkcg_policy_data	*cpd[BLKCG_MAX_POLS];
 
 	struct list_head		all_blkcgs_node;
+
+	/*
+	 * List of updated percpu blkg_iostat_set's since the last flush.
+	 */
+	struct llist_head __percpu	*lhead;
+
 #ifdef CONFIG_BLK_CGROUP_FC_APPID
 	char                            fc_app_id[FC_APPID_LEN];
 #endif
-- 
2.31.1